Road disease real-time identification method based on multi-channel three-dimensional ground penetrating radar

Abstract

The invention discloses a road disease real-time identification method based on a multi-channel three-dimensional ground penetrating radar, and relates to the technical field of road disease identification, and the method comprises the steps: S1, scanning a to-be-detected road through the three-dimensional ground penetrating radar, and selecting a plurality of channels from a plurality of channels of the three-dimensional ground penetrating radar for data extraction; s2, performing feature fusion on the extracted channel image data to obtain a first fusion feature map; s3, constructing and training a similarity calculation network; s4, inputting the first fusion feature map into the trained similarity calculation network, and carrying out disease recognition; and S5, if the disease is identified, calling the two-dimensional radar to perform re-check confirmation at the position corresponding to the disease. According to the method, the similarity network is utilized, the real-time weak features can correspond to the post-processing strong features, so that the diseases are effectively judged, partial channels of the three-dimensional radar are utilized to replace full-channel post-processing for real-time feature fusion, the problem that the real-time data features are weak is solved, and real-time images can be used for disease judgment.

Description

Technical Field

[0001] This invention relates to the field of road defect identification technology, specifically to a real-time road defect identification method based on multi-channel three-dimensional ground-penetrating radar. Background Technology

[0002] Three-dimensional ground-penetrating radar (3D GPR) has the ability to simultaneously acquire underground reflected signals through multiple channels. Current technologies for identifying road defects using 3D GPR typically include: multi-channel data acquisition, post-processing channel fusion, image enhancement, and then combining this with manual or algorithmic identification. While this significantly improves the spatial coverage and data integrity of municipal road defect detection, it has problems in the following aspects: 1) Poor real-time performance: Although 3D ground-penetrating radar has multi-channel acquisition capabilities, the fusion of multi-channel data and imaging processing require a large amount of computation, making it impossible to provide a disease diagnosis immediately after scanning. It usually requires a long post-processing time, which affects the efficiency of the operation.

[0003] 2) The disease characteristics are not obvious. The real-time images of 3D radar are mostly single-channel or few-channel data. Affected by noise and reflection distortion, the characteristics are not as obvious as the images after post-processing and fusion, and it is difficult to use them directly for disease identification.

[0004] 3) Two-dimensional radar verification is still required. Even if three-dimensional radar identifies suspected defects after post-processing, two-dimensional radar must still be used for verification to increase the reliability of the identification results. However, if the initial judgment from three-dimensional radar cannot be output in real time, it will cause unnecessary waiting in the workflow.

[0005] 4) There is a lack of algorithmic frameworks that can effectively utilize real-time data. Currently, there is no mature method that can directly utilize the real-time, low-channel data of 3D radar for effective feature fusion and establish a reliable correspondence with the high-quality images after post-processing, thereby achieving accurate real-time disease identification.

[0006] In summary, existing technologies cannot determine the presence of defects in municipal roads immediately after scanning, nor can they directly guide the verification work of 2D radar based on real-time results, thus affecting the overall detection efficiency and the speed of on-site decision-making. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for real-time identification of road defects based on multi-channel three-dimensional ground-penetrating radar.

[0008] The objective of this invention is achieved through the following technical solution: This application discloses a method for real-time identification of road defects based on multi-channel three-dimensional ground-penetrating radar, including the following steps: S1. Scan the road to be detected using a 3D ground-penetrating radar, and select several channels from the multiple channels of the 3D ground-penetrating radar for data extraction. S2. Perform feature fusion on the channel image data extracted in step S1 to obtain the first fused feature map; S3. Construct a similarity calculation network and train it; S4. Input the first fusion feature map obtained in step S2 into the trained similarity calculation network to identify diseases; S5. If a defect is identified, the location of the defect will be verified on-site by using a two-dimensional radar.

[0009] Furthermore, the feature fusion described in step S2 includes weighted fusion, spatiotemporal correlation convolution, or feature overlay.

[0010] Furthermore, step S3 specifically includes the following steps: S31. After post-processing, a second fused feature map is obtained, and it is manually labeled to indicate whether there are any defects. S32. Construct a similarity calculation network to learn the relationship between the first fused feature map and the second fused feature map; S33. Using the first fused feature map and the second fused feature map as training samples, train the similarity calculation network constructed in step S32 and output the similarity score.

[0011] Preferably, the architecture of the similarity calculation network includes a dual-branch convolutional encoder, a feature normalization layer, and a similarity calculation module. The dual-branch convolutional encoder receives a first fused feature map and a second fused feature map respectively, and extracts the corresponding deep embedding features through the feature encoder. The feature normalization layer normalizes the extracted deep embedding features. Based on the normalized deep embedding features, the similarity score between the first fused feature map and the second fused feature map is calculated through the similarity calculation module.

[0012] Preferably, step S3 specifically includes the following steps: S31, via coding network For the first fused feature map The process is performed to obtain the first embedded feature vector. , Through coding network For the second fusion feature map The process is performed to obtain the second embedded feature vector. , ; S32, For the first embedded feature vector Perform L2 normalization to obtain the normalized first embedded feature vector. , ,in Represents the first embedded feature vector The L2 norm of the second embedded feature vector Perform L2 normalization to obtain the normalized second embedded feature vector. , ,in Represents the second embedded feature vector The L2 norm; S33. The normalized first embedded feature vector from step S32 is... and the normalized second embedded feature vector Perform splicing to construct differential features , ; S34, Based on the constructed differential features To achieve a learnable mapping for similarity scores: ,in Indicates the similarity score. This represents the sigmoid function. and They represent the learnable parameters, Indicates the activation function; S35, Based on Contrast Loss The similarity calculation module is trained. ,in This represents the Euclidean distance, where m represents the hyperparameter. Indicates a label, if ,but For a matching pair, if ,but These are non-matching pairs.

[0013] Preferably, step S4 specifically includes: fusing the first fused feature map The input is fed into the trained similarity calculation network to calculate its similarity with the second fused feature map. The similarity score is calculated, and if the calculated similarity score exceeds a preset threshold, then the first fused feature map... The presence of disease affects the second fusion feature map. The labels are used to vote, and the final disease category and confidence level are obtained.

[0014] The beneficial effects of this invention are: 1) This application utilizes a portion of the channels of a three-dimensional radar to replace the post-processing of the entire channel, performs real-time feature fusion, eliminates time-consuming post-processing steps, and can provide a preliminary disease judgment immediately upon completion of scanning, greatly improving the real-time performance of detection, solving the problem of weak real-time data features, and enabling real-time images to be used for disease judgment.

[0015] 2) In this application, the neural network used to calculate similarity uses some key channel images as input, so that real-time weak features can be mapped to post-processed strong features, thereby effectively judging diseases, reducing network overhead, and providing the possibility of building a lightweight network for deployment on edge computing devices to achieve real-time disease identification.

[0016] 3) Optimize the 2D radar verification process by integrating the real-time identification results of 3D radar with the 2D radar verification process. When a suspected defect area is identified, 2D radar verification can be performed immediately, avoiding post-processing waiting time. The initial scan and verification work can be completed in one operation, which will likely reduce workload and operating costs, and realize the simplification, intelligence and efficiency of the detection process.

[0017] 4) Improve the intelligence level of disease identification. By learning from a large amount of post-processed labeled data, the real-time identification capability is continuously enhanced, and it has self-learning and scalability. At the same time, the identification stability is improved by voting on multiple similar samples, avoiding single judgment errors. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the steps of a real-time road defect identification method based on multi-channel three-dimensional ground-penetrating radar according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This application discloses a real-time road defect identification method based on multi-channel 3D ground-penetrating radar. Utilizing real-time images from partial channels acquired by 3D ground-penetrating radar, and through feature fusion and similarity calculation networks, it achieves real-time identification of potential defects. Based on the judgment results, it automatically triggers 2D radar verification, thereby avoiding the time delay caused by traditional full-channel post-processing. A schematic diagram of the method's steps is shown below. Figure 1 As shown, the specific steps include: S1. Scan the road to be detected using a 3D ground-penetrating radar, and select several channels from the multiple channels of the 3D ground-penetrating radar for data extraction. S2. Perform feature fusion on the channel image data extracted in step S1 to obtain the first fused feature map (real-time image fusion feature map); the feature fusion includes weighted fusion, spatiotemporal correlation convolution or feature overlay, etc. S3. Construct a similarity calculation network and train it; S4. Input the first fusion feature map obtained in step S2 into the trained similarity calculation network to identify diseases; S5. If a defect is identified, the location of the defect will be verified on-site by using a two-dimensional radar.

[0021] Specifically, step S3 includes the following steps: S31. After post-processing, the second fusion feature map (post-processed fusion feature map) is obtained offline, and manually labeled to indicate whether there are diseases. S32. Construct a similarity calculation network to learn the relationship between the first fused feature map and the second fused feature map; S33. Using the first fused feature map and the second fused feature map as training samples, train the similarity calculation network constructed in step S32 and output the similarity score.

[0022] Specifically, the architecture of the similarity calculation network includes a dual-branch convolutional encoder (Siamese Encoder), a feature normalization layer, and a similarity calculation module, used to establish a metric space relationship between the first fused feature map and the second fused feature map. The similarity calculation network is a dual-branch feature extraction structure. The dual-branch convolutional encoder receives the first fused feature map and the second fused feature map respectively, and extracts the corresponding deep embedding features through the feature encoder. The feature normalization layer normalizes the extracted deep embedding features. Based on the normalized deep embedding features, the similarity score between the first fused feature map and the second fused feature map is calculated through the similarity calculation module.

[0023] Specifically, the similarity calculation network treats matching first and second fused feature maps as a matching pair and labels them with 1, while treating non-matching first and second fused feature maps as a matching pair and labeling them with 0. During training, the network prioritizes the similarity of matching pairs, and the similarity of non-matching pairs is determined by the feature differences between the input feature map pairs (achieved by constructing the difference features and introducing a distance penalty for feature map pairs in the loss function). Step S3 specifically includes the following steps: S31, via coding network For the first fused feature map The process is performed to obtain the first embedded feature vector. , Through coding network For the second fusion feature map The process is performed to obtain the second embedded feature vector. , First fusion feature map Second fusion feature map The coding network used With the same structure and shared parameters, the encoding network can be composed of convolutional layers, BatchNorm layers and ReLU activation layers, or it can adopt a lightweight network (such as MobileNet, ShuffleNet) or a residual structure (such as ResNet Block). S32, For the first embedded feature vector Perform L2 normalization to obtain the normalized first embedded feature vector. , ,in Represents the first embedded feature vector The L2 norm of the second embedded feature vector Perform L2 normalization to obtain the normalized second embedded feature vector. , ,in Represents the second embedded feature vector The L2 norm; S33. The normalized first embedded feature vector from step S32 is... and the normalized second embedded feature vector Perform splicing to construct differential features , ; S34, Based on the constructed differential features To achieve a learnable mapping for similarity scores: ,in Indicates the similarity score. This represents the sigmoid function. and They represent the learnable parameters, Indicates the activation function; S35. To ensure that the first fused feature map and its true corresponding second fused feature map achieve the highest similarity during training, this invention employs a supervised loss based on metric learning, namely, contrastive loss. The similarity calculation module is trained. ,in Euclidean distance is expressed by the following formula: 'm' represents a hyperparameter used to force unmatched pairs to maintain a large distance. Indicates a label, if ,but For a matching pair, if ,but These are non-matching pairs. The loss function ensures that the similarity score of the true matching pairs is maximized, enabling the network to learn the mapping relationship between "real-time partial channel features → post-processing features".

[0024] Specifically, step S4 includes: fusing the first fused feature map The input is fed into the trained similarity calculation network to calculate its similarity with the second fused feature map. The similarity score is calculated, and if the calculated similarity score exceeds a preset threshold, then the first fused feature map... The presence of disease affects the second fusion feature map. The system uses tags to vote, resulting in the final disease category and confidence level. By deploying this neural network on an Nvidia Jetson Nano edge computing device and integrating it with a ground-penetrating radar real-time acquisition device, real-time disease identification can be achieved.

[0025] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for real-time identification of road defects based on multi-channel three-dimensional ground-penetrating radar, characterized in that, Includes the following steps: S1. Scan the road to be detected using a 3D ground-penetrating radar, and select several channels from the multiple channels of the 3D ground-penetrating radar for data extraction. S2. Perform feature fusion on the channel image data extracted in step S1 to obtain the first fused feature map; S3. Construct a similarity calculation network and train it; S4. Input the first fusion feature map obtained in step S2 into the trained similarity calculation network to identify diseases; S5. If a defect is identified, the location of the defect will be verified on-site by using a two-dimensional radar.

2. The method for real-time identification of road defects based on multi-channel three-dimensional ground-penetrating radar according to claim 1, characterized in that: The feature fusion described in step S2 includes weighted fusion, spatiotemporal correlation convolution, or feature overlay.

3. The method for real-time identification of road defects based on multi-channel three-dimensional ground-penetrating radar according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. After post-processing, a second fused feature map is obtained, and it is manually labeled to indicate whether there are any defects. S32. Construct a similarity calculation network to learn the relationship between the first fused feature map and the second fused feature map; S33. Using the first fused feature map and the second fused feature map as training samples, train the similarity calculation network constructed in step S32 and output the similarity score.

4. The method for real-time identification of road defects based on multi-channel three-dimensional ground-penetrating radar according to claim 3, characterized in that: The architecture of the similarity calculation network includes a dual-branch convolutional encoder, a feature normalization layer, and a similarity calculation module. The dual-branch convolutional encoder receives a first fused feature map and a second fused feature map respectively, and extracts the corresponding deep embedding features through the feature encoder. The feature normalization layer normalizes the extracted deep embedding features; Based on the normalized deep embedding features, the similarity score between the first fused feature map and the second fused feature map is calculated through the similarity calculation module.

5. The method for real-time identification of road defects based on multi-channel three-dimensional ground-penetrating radar according to claim 4, characterized in that, Step S3 specifically includes the following steps: S31, via coding network For the first fused feature map The process is performed to obtain the first embedded feature vector. , Through coding network For the second fusion feature map The process is performed to obtain the second embedded feature vector. , ; S32, For the first embedded feature vector Perform L2 normalization to obtain the normalized first embedded feature vector. , ,in Represents the first embedded feature vector The L2 norm of the second embedded feature vector Perform L2 normalization to obtain the normalized second embedded feature vector. , ,in Represents the second embedded feature vector The L2 norm; S33. The normalized first embedded feature vector from step S32 is... and the normalized second embedded feature vector Perform splicing to construct differential features , ; S34, Based on the constructed differential features To achieve a learnable mapping for similarity scores: ,in Indicates the similarity score. This represents the sigmoid function. and They represent the learnable parameters, Indicates the activation function; S35, Based on Contrast Loss The similarity calculation module is trained. ,in This represents the Euclidean distance, where m represents the hyperparameter. Indicates a label, if ,but For a matching pair, if ,but These are non-matching pairs.

6. The method for real-time identification of road defects based on multi-channel three-dimensional ground-penetrating radar according to claim 5, characterized in that, Step S4 specifically includes: fusing the first fused feature map The input is fed into the trained similarity calculation network to calculate its similarity with the second fused feature map. The similarity score is calculated, and if the calculated similarity score exceeds a preset threshold, then the first fused feature map... The presence of disease affects the second fusion feature map. The labels are used to vote, and the final disease category and confidence level are obtained.

Images